les: implement client connection logic (#16899)

This PR implements les.freeClientPool. It also adds a simulated clock
in common/mclock, which enables time-sensitive tests to run quickly
and still produce accurate results, and package common/prque which is
a generalised variant of prque that enables removing elements other
than the top one from the queue.

les.freeClientPool implements a client database that limits the
connection time of each client and manages accepting/rejecting
incoming connections and even kicking out some connected clients. The
pool calculates recent usage time for each known client (a value that
increases linearly when the client is connected and decreases
exponentially when not connected). Clients with lower recent usage are
preferred, unknown nodes have the highest priority. Already connected
nodes receive a small bias in their favor in order to avoid accepting
and instantly kicking out clients.

Note: the pool can use any string for client identification. Using
signature keys for that purpose would not make sense when being known
has a negative value for the client. Currently the LES protocol
manager uses IP addresses (without port address) to identify clients.

7 files changed, 761 insertions, 1 deletions

diff --git a/common/mclock/mclock.go b/common/mclock/mclock.go
index 02608d17b..dcac59c6c 100644
--- a/common/mclock/mclock.go
+++ b/common/mclock/mclock.go
@@ -30,3 +30,34 @@ type AbsTime time.Duration
 func Now() AbsTime {
 	return AbsTime(monotime.Now())
 }
+
+// Add returns t + d.
+func (t AbsTime) Add(d time.Duration) AbsTime {
+	return t + AbsTime(d)
+}
+
+// Clock interface makes it possible to replace the monotonic system clock with
+// a simulated clock.
+type Clock interface {
+	Now() AbsTime
+	Sleep(time.Duration)
+	After(time.Duration) <-chan time.Time
+}
+
+// System implements Clock using the system clock.
+type System struct{}
+
+// Now implements Clock.
+func (System) Now() AbsTime {
+	return AbsTime(monotime.Now())
+}
+
+// Sleep implements Clock.
+func (System) Sleep(d time.Duration) {
+	time.Sleep(d)
+}
+
+// After implements Clock.
+func (System) After(d time.Duration) <-chan time.Time {
+	return time.After(d)
+}
diff --git a/common/mclock/simclock.go b/common/mclock/simclock.go
new file mode 100644
index 000000000..e014f5615
--- /dev/null
+++ b/common/mclock/simclock.go
@@ -0,0 +1,129 @@
+// Copyright 2018 The go-ethereum Authors
+// This file is part of the go-ethereum library.
+//
+// The go-ethereum library is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Lesser General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// The go-ethereum library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Lesser General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public License
+// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
+
+package mclock
+
+import (
+	"sync"
+	"time"
+)
+
+// Simulated implements a virtual Clock for reproducible time-sensitive tests. It
+// simulates a scheduler on a virtual timescale where actual processing takes zero time.
+//
+// The virtual clock doesn't advance on its own, call Run to advance it and execute timers.
+// Since there is no way to influence the Go scheduler, testing timeout behaviour involving
+// goroutines needs special care. A good way to test such timeouts is as follows: First
+// perform the action that is supposed to time out. Ensure that the timer you want to test
+// is created. Then run the clock until after the timeout. Finally observe the effect of
+// the timeout using a channel or semaphore.
+type Simulated struct {
+	now       AbsTime
+	scheduled []event
+	mu        sync.RWMutex
+	cond      *sync.Cond
+}
+
+type event struct {
+	do func()
+	at AbsTime
+}
+
+// Run moves the clock by the given duration, executing all timers before that duration.
+func (s *Simulated) Run(d time.Duration) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.init()
+
+	end := s.now + AbsTime(d)
+	for len(s.scheduled) > 0 {
+		ev := s.scheduled[0]
+		if ev.at > end {
+			break
+		}
+		s.now = ev.at
+		ev.do()
+		s.scheduled = s.scheduled[1:]
+	}
+	s.now = end
+}
+
+func (s *Simulated) ActiveTimers() int {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+
+	return len(s.scheduled)
+}
+
+func (s *Simulated) WaitForTimers(n int) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.init()
+
+	for len(s.scheduled) < n {
+		s.cond.Wait()
+	}
+}
+
+// Now implements Clock.
+func (s *Simulated) Now() AbsTime {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+
+	return s.now
+}
+
+// Sleep implements Clock.
+func (s *Simulated) Sleep(d time.Duration) {
+	<-s.After(d)
+}
+
+// After implements Clock.
+func (s *Simulated) After(d time.Duration) <-chan time.Time {
+	after := make(chan time.Time, 1)
+	s.insert(d, func() {
+		after <- (time.Time{}).Add(time.Duration(s.now))
+	})
+	return after
+}
+
+func (s *Simulated) insert(d time.Duration, do func()) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.init()
+
+	at := s.now + AbsTime(d)
+	l, h := 0, len(s.scheduled)
+	ll := h
+	for l != h {
+		m := (l + h) / 2
+		if at < s.scheduled[m].at {
+			h = m
+		} else {
+			l = m + 1
+		}
+	}
+	s.scheduled = append(s.scheduled, event{})
+	copy(s.scheduled[l+1:], s.scheduled[l:ll])
+	s.scheduled[l] = event{do: do, at: at}
+	s.cond.Broadcast()
+}
+
+func (s *Simulated) init() {
+	if s.cond == nil {
+		s.cond = sync.NewCond(&s.mu)
+	}
+}
diff --git a/common/prque/prque.go b/common/prque/prque.go
new file mode 100755
index 000000000..9fd31a2e5
--- /dev/null
+++ b/common/prque/prque.go
@@ -0,0 +1,57 @@
+// This is a duplicated and slightly modified version of "gopkg.in/karalabe/cookiejar.v2/collections/prque".
+
+package prque
+
+import (
+	"container/heap"
+)
+
+// Priority queue data structure.
+type Prque struct {
+	cont *sstack
+}
+
+// Creates a new priority queue.
+func New(setIndex setIndexCallback) *Prque {
+	return &Prque{newSstack(setIndex)}
+}
+
+// Pushes a value with a given priority into the queue, expanding if necessary.
+func (p *Prque) Push(data interface{}, priority int64) {
+	heap.Push(p.cont, &item{data, priority})
+}
+
+// Pops the value with the greates priority off the stack and returns it.
+// Currently no shrinking is done.
+func (p *Prque) Pop() (interface{}, int64) {
+	item := heap.Pop(p.cont).(*item)
+	return item.value, item.priority
+}
+
+// Pops only the item from the queue, dropping the associated priority value.
+func (p *Prque) PopItem() interface{} {
+	return heap.Pop(p.cont).(*item).value
+}
+
+// Remove removes the element with the given index.
+func (p *Prque) Remove(i int) interface{} {
+	if i < 0 {
+		return nil
+	}
+	return heap.Remove(p.cont, i)
+}
+
+// Checks whether the priority queue is empty.
+func (p *Prque) Empty() bool {
+	return p.cont.Len() == 0
+}
+
+// Returns the number of element in the priority queue.
+func (p *Prque) Size() int {
+	return p.cont.Len()
+}
+
+// Clears the contents of the priority queue.
+func (p *Prque) Reset() {
+	*p = *New(p.cont.setIndex)
+}
diff --git a/common/prque/sstack.go b/common/prque/sstack.go
new file mode 100755
index 000000000..4875dae99
--- /dev/null
+++ b/common/prque/sstack.go
@@ -0,0 +1,106 @@
+// This is a duplicated and slightly modified version of "gopkg.in/karalabe/cookiejar.v2/collections/prque".
+
+package prque
+
+// The size of a block of data
+const blockSize = 4096
+
+// A prioritized item in the sorted stack.
+//
+// Note: priorities can "wrap around" the int64 range, a comes before b if (a.priority - b.priority) > 0.
+// The difference between the lowest and highest priorities in the queue at any point should be less than 2^63.
+type item struct {
+	value    interface{}
+	priority int64
+}
+
+// setIndexCallback is called when the element is moved to a new index.
+// Providing setIndexCallback is optional, it is needed only if the application needs
+// to delete elements other than the top one.
+type setIndexCallback func(a interface{}, i int)
+
+// Internal sortable stack data structure. Implements the Push and Pop ops for
+// the stack (heap) functionality and the Len, Less and Swap methods for the
+// sortability requirements of the heaps.
+type sstack struct {
+	setIndex setIndexCallback
+	size     int
+	capacity int
+	offset   int
+
+	blocks [][]*item
+	active []*item
+}
+
+// Creates a new, empty stack.
+func newSstack(setIndex setIndexCallback) *sstack {
+	result := new(sstack)
+	result.setIndex = setIndex
+	result.active = make([]*item, blockSize)
+	result.blocks = [][]*item{result.active}
+	result.capacity = blockSize
+	return result
+}
+
+// Pushes a value onto the stack, expanding it if necessary. Required by
+// heap.Interface.
+func (s *sstack) Push(data interface{}) {
+	if s.size == s.capacity {
+		s.active = make([]*item, blockSize)
+		s.blocks = append(s.blocks, s.active)
+		s.capacity += blockSize
+		s.offset = 0
+	} else if s.offset == blockSize {
+		s.active = s.blocks[s.size/blockSize]
+		s.offset = 0
+	}
+	if s.setIndex != nil {
+		s.setIndex(data.(*item).value, s.size)
+	}
+	s.active[s.offset] = data.(*item)
+	s.offset++
+	s.size++
+}
+
+// Pops a value off the stack and returns it. Currently no shrinking is done.
+// Required by heap.Interface.
+func (s *sstack) Pop() (res interface{}) {
+	s.size--
+	s.offset--
+	if s.offset < 0 {
+		s.offset = blockSize - 1
+		s.active = s.blocks[s.size/blockSize]
+	}
+	res, s.active[s.offset] = s.active[s.offset], nil
+	if s.setIndex != nil {
+		s.setIndex(res.(*item).value, -1)
+	}
+	return
+}
+
+// Returns the length of the stack. Required by sort.Interface.
+func (s *sstack) Len() int {
+	return s.size
+}
+
+// Compares the priority of two elements of the stack (higher is first).
+// Required by sort.Interface.
+func (s *sstack) Less(i, j int) bool {
+	return (s.blocks[i/blockSize][i%blockSize].priority - s.blocks[j/blockSize][j%blockSize].priority) > 0
+}
+
+// Swaps two elements in the stack. Required by sort.Interface.
+func (s *sstack) Swap(i, j int) {
+	ib, io, jb, jo := i/blockSize, i%blockSize, j/blockSize, j%blockSize
+	a, b := s.blocks[jb][jo], s.blocks[ib][io]
+	if s.setIndex != nil {
+		s.setIndex(a.value, i)
+		s.setIndex(b.value, j)
+	}
+	s.blocks[ib][io], s.blocks[jb][jo] = a, b
+}
+
+// Resets the stack, effectively clearing its contents.
+func (s *sstack) Reset() {
+	*s = *newSstack(s.setIndex)
+}
diff --git a/les/freeclient.go b/les/freeclient.go
new file mode 100644
index 000000000..5ee607be8
--- /dev/null
+++ b/les/freeclient.go
@@ -0,0 +1,278 @@
+// Copyright 2016 The go-ethereum Authors
+// This file is part of the go-ethereum library.
+//
+// The go-ethereum library is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Lesser General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// The go-ethereum library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Lesser General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public License
+// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
+
+// Package les implements the Light Ethereum Subprotocol.
+package les
+
+import (
+	"io"
+	"math"
+	"sync"
+	"time"
+
+	"github.com/ethereum/go-ethereum/common/mclock"
+	"github.com/ethereum/go-ethereum/common/prque"
+	"github.com/ethereum/go-ethereum/ethdb"
+	"github.com/ethereum/go-ethereum/log"
+	"github.com/ethereum/go-ethereum/rlp"
+)
+
+// freeClientPool implements a client database that limits the connection time
+// of each client and manages accepting/rejecting incoming connections and even
+// kicking out some connected clients. The pool calculates recent usage time
+// for each known client (a value that increases linearly when the client is
+// connected and decreases exponentially when not connected). Clients with lower
+// recent usage are preferred, unknown nodes have the highest priority. Already
+// connected nodes receive a small bias in their favor in order to avoid accepting
+// and instantly kicking out clients.
+//
+// Note: the pool can use any string for client identification. Using signature
+// keys for that purpose would not make sense when being known has a negative
+// value for the client. Currently the LES protocol manager uses IP addresses
+// (without port address) to identify clients.
+type freeClientPool struct {
+	db     ethdb.Database
+	lock   sync.Mutex
+	clock  mclock.Clock
+	closed bool
+
+	connectedLimit, totalLimit int
+
+	addressMap            map[string]*freeClientPoolEntry
+	connPool, disconnPool *prque.Prque
+	startupTime           mclock.AbsTime
+	logOffsetAtStartup    int64
+}
+
+const (
+	recentUsageExpTC     = time.Hour   // time constant of the exponential weighting window for "recent" server usage
+	fixedPointMultiplier = 0x1000000   // constant to convert logarithms to fixed point format
+	connectedBias        = time.Minute // this bias is applied in favor of already connected clients in order to avoid kicking them out very soon
+)
+
+// newFreeClientPool creates a new free client pool
+func newFreeClientPool(db ethdb.Database, connectedLimit, totalLimit int, clock mclock.Clock) *freeClientPool {
+	pool := &freeClientPool{
+		db:             db,
+		clock:          clock,
+		addressMap:     make(map[string]*freeClientPoolEntry),
+		connPool:       prque.New(poolSetIndex),
+		disconnPool:    prque.New(poolSetIndex),
+		connectedLimit: connectedLimit,
+		totalLimit:     totalLimit,
+	}
+	pool.loadFromDb()
+	return pool
+}
+
+func (f *freeClientPool) stop() {
+	f.lock.Lock()
+	f.closed = true
+	f.saveToDb()
+	f.lock.Unlock()
+}
+
+// connect should be called after a successful handshake. If the connection was
+// rejected, there is no need to call disconnect.
+//
+// Note: the disconnectFn callback should not block.
+func (f *freeClientPool) connect(address string, disconnectFn func()) bool {
+	f.lock.Lock()
+	defer f.lock.Unlock()
+
+	if f.closed {
+		return false
+	}
+	e := f.addressMap[address]
+	now := f.clock.Now()
+	var recentUsage int64
+	if e == nil {
+		e = &freeClientPoolEntry{address: address, index: -1}
+		f.addressMap[address] = e
+	} else {
+		if e.connected {
+			log.Debug("Client already connected", "address", address)
+			return false
+		}
+		recentUsage = int64(math.Exp(float64(e.logUsage-f.logOffset(now)) / fixedPointMultiplier))
+	}
+	e.linUsage = recentUsage - int64(now)
+	// check whether (linUsage+connectedBias) is smaller than the highest entry in the connected pool
+	if f.connPool.Size() == f.connectedLimit {
+		i := f.connPool.PopItem().(*freeClientPoolEntry)
+		if e.linUsage+int64(connectedBias)-i.linUsage < 0 {
+			// kick it out and accept the new client
+			f.connPool.Remove(i.index)
+			f.calcLogUsage(i, now)
+			i.connected = false
+			f.disconnPool.Push(i, -i.logUsage)
+			log.Debug("Client kicked out", "address", i.address)
+			i.disconnectFn()
+		} else {
+			// keep the old client and reject the new one
+			f.connPool.Push(i, i.linUsage)
+			log.Debug("Client rejected", "address", address)
+			return false
+		}
+	}
+	f.disconnPool.Remove(e.index)
+	e.connected = true
+	e.disconnectFn = disconnectFn
+	f.connPool.Push(e, e.linUsage)
+	if f.connPool.Size()+f.disconnPool.Size() > f.totalLimit {
+		f.disconnPool.Pop()
+	}
+	log.Debug("Client accepted", "address", address)
+	return true
+}
+
+// disconnect should be called when a connection is terminated. If the disconnection
+// was initiated by the pool itself using disconnectFn then calling disconnect is
+// not necessary but permitted.
+func (f *freeClientPool) disconnect(address string) {
+	f.lock.Lock()
+	defer f.lock.Unlock()
+
+	if f.closed {
+		return
+	}
+	e := f.addressMap[address]
+	now := f.clock.Now()
+	if !e.connected {
+		log.Debug("Client already disconnected", "address", address)
+		return
+	}
+
+	f.connPool.Remove(e.index)
+	f.calcLogUsage(e, now)
+	e.connected = false
+	f.disconnPool.Push(e, -e.logUsage)
+	log.Debug("Client disconnected", "address", address)
+}
+
+// logOffset calculates the time-dependent offset for the logarithmic
+// representation of recent usage
+func (f *freeClientPool) logOffset(now mclock.AbsTime) int64 {
+	// Note: fixedPointMultiplier acts as a multiplier here; the reason for dividing the divisor
+	// is to avoid int64 overflow. We assume that int64(recentUsageExpTC) >> fixedPointMultiplier.
+	logDecay := int64((time.Duration(now - f.startupTime)) / (recentUsageExpTC / fixedPointMultiplier))
+	return f.logOffsetAtStartup + logDecay
+}
+
+// calcLogUsage converts recent usage from linear to logarithmic representation
+// when disconnecting a peer or closing the client pool
+func (f *freeClientPool) calcLogUsage(e *freeClientPoolEntry, now mclock.AbsTime) {
+	dt := e.linUsage + int64(now)
+	if dt < 1 {
+		dt = 1
+	}
+	e.logUsage = int64(math.Log(float64(dt))*fixedPointMultiplier) + f.logOffset(now)
+}
+
+// freeClientPoolStorage is the RLP representation of the pool's database storage
+type freeClientPoolStorage struct {
+	LogOffset uint64
+	List      []*freeClientPoolEntry
+}
+
+// loadFromDb restores pool status from the database storage
+// (automatically called at initialization)
+func (f *freeClientPool) loadFromDb() {
+	enc, err := f.db.Get([]byte("freeClientPool"))
+	if err != nil {
+		return
+	}
+	var storage freeClientPoolStorage
+	err = rlp.DecodeBytes(enc, &storage)
+	if err != nil {
+		log.Error("Failed to decode client list", "err", err)
+		return
+	}
+	f.logOffsetAtStartup = int64(storage.LogOffset)
+	f.startupTime = f.clock.Now()
+	for _, e := range storage.List {
+		log.Debug("Loaded free client record", "address", e.address, "logUsage", e.logUsage)
+		f.addressMap[e.address] = e
+		f.disconnPool.Push(e, -e.logUsage)
+	}
+}
+
+// saveToDb saves pool status to the database storage
+// (automatically called during shutdown)
+func (f *freeClientPool) saveToDb() {
+	now := f.clock.Now()
+	storage := freeClientPoolStorage{
+		LogOffset: uint64(f.logOffset(now)),
+		List:      make([]*freeClientPoolEntry, len(f.addressMap)),
+	}
+	i := 0
+	for _, e := range f.addressMap {
+		if e.connected {
+			f.calcLogUsage(e, now)
+		}
+		storage.List[i] = e
+		i++
+	}
+	enc, err := rlp.EncodeToBytes(storage)
+	if err != nil {
+		log.Error("Failed to encode client list", "err", err)
+	} else {
+		f.db.Put([]byte("freeClientPool"), enc)
+	}
+}
+
+// freeClientPoolEntry represents a client address known by the pool.
+// When connected, recent usage is calculated as linUsage + int64(clock.Now())
+// When disconnected, it is calculated as exp(logUsage - logOffset) where logOffset
+// also grows linearly with time while the server is running.
+// Conversion between linear and logarithmic representation happens when connecting
+// or disconnecting the node.
+//
+// Note: linUsage and logUsage are values used with constantly growing offsets so
+// even though they are close to each other at any time they may wrap around int64
+// limits over time. Comparison should be performed accordingly.
+type freeClientPoolEntry struct {
+	address            string
+	connected          bool
+	disconnectFn       func()
+	linUsage, logUsage int64
+	index              int
+}
+
+func (e *freeClientPoolEntry) EncodeRLP(w io.Writer) error {
+	return rlp.Encode(w, []interface{}{e.address, uint64(e.logUsage)})
+}
+
+func (e *freeClientPoolEntry) DecodeRLP(s *rlp.Stream) error {
+	var entry struct {
+		Address  string
+		LogUsage uint64
+	}
+	if err := s.Decode(&entry); err != nil {
+		return err
+	}
+	e.address = entry.Address
+	e.logUsage = int64(entry.LogUsage)
+	e.connected = false
+	e.index = -1
+	return nil
+}
+
+// poolSetIndex callback is used by both priority queues to set/update the index of
+// the element in the queue. Index is needed to remove elements other than the top one.
+func poolSetIndex(a interface{}, i int) {
+	a.(*freeClientPoolEntry).index = i
+}
diff --git a/les/freeclient_test.go b/les/freeclient_test.go
new file mode 100644
index 000000000..e95abc7aa
--- /dev/null
+++ b/les/freeclient_test.go
@@ -0,0 +1,139 @@
+// Copyright 2017 The go-ethereum Authors
+// This file is part of the go-ethereum library.
+//
+// The go-ethereum library is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Lesser General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// The go-ethereum library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Lesser General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public License
+// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
+
+// Package light implements on-demand retrieval capable state and chain objects
+// for the Ethereum Light Client.
+package les
+
+import (
+	"fmt"
+	"math/rand"
+	"testing"
+	"time"
+
+	"github.com/ethereum/go-ethereum/common/mclock"
+	"github.com/ethereum/go-ethereum/ethdb"
+)
+
+func TestFreeClientPoolL10C100(t *testing.T) {
+	testFreeClientPool(t, 10, 100)
+}
+
+func TestFreeClientPoolL40C200(t *testing.T) {
+	testFreeClientPool(t, 40, 200)
+}
+
+func TestFreeClientPoolL100C300(t *testing.T) {
+	testFreeClientPool(t, 100, 300)
+}
+
+const testFreeClientPoolTicks = 500000
+
+func testFreeClientPool(t *testing.T, connLimit, clientCount int) {
+	var (
+		clock     mclock.Simulated
+		db        = ethdb.NewMemDatabase()
+		pool      = newFreeClientPool(db, connLimit, 10000, &clock)
+		connected = make([]bool, clientCount)
+		connTicks = make([]int, clientCount)
+		disconnCh = make(chan int, clientCount)
+	)
+	peerId := func(i int) string {
+		return fmt.Sprintf("test peer #%d", i)
+	}
+	disconnFn := func(i int) func() {
+		return func() {
+			disconnCh <- i
+		}
+	}
+
+	// pool should accept new peers up to its connected limit
+	for i := 0; i < connLimit; i++ {
+		if pool.connect(peerId(i), disconnFn(i)) {
+			connected[i] = true
+		} else {
+			t.Fatalf("Test peer #%d rejected", i)
+		}
+	}
+	// since all accepted peers are new and should not be kicked out, the next one should be rejected
+	if pool.connect(peerId(connLimit), disconnFn(connLimit)) {
+		connected[connLimit] = true
+		t.Fatalf("Peer accepted over connected limit")
+	}
+
+	// randomly connect and disconnect peers, expect to have a similar total connection time at the end
+	for tickCounter := 0; tickCounter < testFreeClientPoolTicks; tickCounter++ {
+		clock.Run(1 * time.Second)
+
+		i := rand.Intn(clientCount)
+		if connected[i] {
+			pool.disconnect(peerId(i))
+			connected[i] = false
+			connTicks[i] += tickCounter
+		} else {
+			if pool.connect(peerId(i), disconnFn(i)) {
+				connected[i] = true
+				connTicks[i] -= tickCounter
+			}
+		}
+	pollDisconnects:
+		for {
+			select {
+			case i := <-disconnCh:
+				pool.disconnect(peerId(i))
+				if connected[i] {
+					connTicks[i] += tickCounter
+					connected[i] = false
+				}
+			default:
+				break pollDisconnects
+			}
+		}
+	}
+
+	expTicks := testFreeClientPoolTicks * connLimit / clientCount
+	expMin := expTicks - expTicks/10
+	expMax := expTicks + expTicks/10
+
+	// check if the total connected time of peers are all in the expected range
+	for i, c := range connected {
+		if c {
+			connTicks[i] += testFreeClientPoolTicks
+		}
+		if connTicks[i] < expMin || connTicks[i] > expMax {
+			t.Errorf("Total connected time of test node #%d (%d) outside expected range (%d to %d)", i, connTicks[i], expMin, expMax)
+		}
+	}
+
+	// a previously unknown peer should be accepted now
+	if !pool.connect("newPeer", func() {}) {
+		t.Fatalf("Previously unknown peer rejected")
+	}
+
+	// close and restart pool
+	pool.stop()
+	pool = newFreeClientPool(db, connLimit, 10000, &clock)
+
+	// try connecting all known peers (connLimit should be filled up)
+	for i := 0; i < clientCount; i++ {
+		pool.connect(peerId(i), func() {})
+	}
+	// expect pool to remember known nodes and kick out one of them to accept a new one
+	if !pool.connect("newPeer2", func() {}) {
+		t.Errorf("Previously unknown peer rejected after restarting pool")
+	}
+	pool.stop()
+}
diff --git a/les/handler.go b/les/handler.go
index 2fc4cde34..91a235bf0 100644
--- a/les/handler.go
+++ b/les/handler.go
@@ -28,6 +28,7 @@ import (
 	"time"
 
 	"github.com/ethereum/go-ethereum/common"
+	"github.com/ethereum/go-ethereum/common/mclock"
 	"github.com/ethereum/go-ethereum/consensus"
 	"github.com/ethereum/go-ethereum/core"
 	"github.com/ethereum/go-ethereum/core/rawdb"
@@ -104,6 +105,7 @@ type ProtocolManager struct {
 	odr         *LesOdr
 	server      *LesServer
 	serverPool  *serverPool
+	clientPool  *freeClientPool
 	lesTopic    discv5.Topic
 	reqDist     *requestDistributor
 	retriever   *retrieveManager
@@ -226,6 +228,7 @@ func (pm *ProtocolManager) Start(maxPeers int) {
 	if pm.lightSync {
 		go pm.syncer()
 	} else {
+		pm.clientPool = newFreeClientPool(pm.chainDb, maxPeers, 10000, mclock.System{})
 		go func() {
 			for range pm.newPeerCh {
 			}
@@ -243,6 +246,9 @@ func (pm *ProtocolManager) Stop() {
 	pm.noMorePeers <- struct{}{}
 
 	close(pm.quitSync) // quits syncer, fetcher
+	if pm.clientPool != nil {
+		pm.clientPool.stop()
+	}
 
 	// Disconnect existing sessions.
 	// This also closes the gate for any new registrations on the peer set.
@@ -264,7 +270,8 @@ func (pm *ProtocolManager) newPeer(pv int, nv uint64, p *p2p.Peer, rw p2p.MsgRea
 // this function terminates, the peer is disconnected.
 func (pm *ProtocolManager) handle(p *peer) error {
 	// Ignore maxPeers if this is a trusted peer
-	if pm.peers.Len() >= pm.maxPeers && !p.Peer.Info().Network.Trusted {
+	// In server mode we try to check into the client pool after handshake
+	if pm.lightSync && pm.peers.Len() >= pm.maxPeers && !p.Peer.Info().Network.Trusted {
 		return p2p.DiscTooManyPeers
 	}
 
@@ -282,6 +289,19 @@ func (pm *ProtocolManager) handle(p *peer) error {
 		p.Log().Debug("Light Ethereum handshake failed", "err", err)
 		return err
 	}
+
+	if !pm.lightSync && !p.Peer.Info().Network.Trusted {
+		addr, ok := p.RemoteAddr().(*net.TCPAddr)
+		// test peer address is not a tcp address, don't use client pool if can not typecast
+		if ok {
+			id := addr.IP.String()
+			if !pm.clientPool.connect(id, func() { go pm.removePeer(p.id) }) {
+				return p2p.DiscTooManyPeers
+			}
+			defer pm.clientPool.disconnect(id)
+		}
+	}
+
 	if rw, ok := p.rw.(*meteredMsgReadWriter); ok {
 		rw.Init(p.version)
 	}